Publication number | US7209056 B2 |

Publication type | Grant |

Application number | US 10/558,710 |

PCT number | PCT/JP2004/012646 |

Publication date | Apr 24, 2007 |

Filing date | Sep 1, 2004 |

Priority date | Sep 2, 2003 |

Fee status | Paid |

Also published as | CN1799200A, CN100495930C, CN101494460A, CN101494460B, EP1662667A1, EP1662667A4, US7317407, US20070018860, US20070139235, WO2005025072A1 |

Publication number | 10558710, 558710, PCT/2004/12646, PCT/JP/2004/012646, PCT/JP/2004/12646, PCT/JP/4/012646, PCT/JP/4/12646, PCT/JP2004/012646, PCT/JP2004/12646, PCT/JP2004012646, PCT/JP200412646, PCT/JP4/012646, PCT/JP4/12646, PCT/JP4012646, PCT/JP412646, US 7209056 B2, US 7209056B2, US-B2-7209056, US7209056 B2, US7209056B2 |

Inventors | Takehiro Moriya, Dai Yang, Noboru Harada |

Original Assignee | Nippon Telegraph And Telephone Corporation |

Export Citation | BiBTeX, EndNote, RefMan |

Patent Citations (5), Non-Patent Citations (3), Referenced by (12), Classifications (21), Legal Events (3) | |

External Links: USPTO, USPTO Assignment, Espacenet | |

US 7209056 B2

Abstract

The average amplitude of the samples in a signal sample sequence X in a floating-point format is determined for each frame. If the average amplitude is greater than a predetermined value, an integer formatting part **12** converts the sequence X into a signal sample sequence Y in a 16-bit integer format by truncation, a compressing part **13** codes the sequence Y to output a code sequence Ca, a difference producing part **14** produces a difference signal Z that corresponds to the difference between the signal sample sequence X and a sequence Y′ in the floating-point format converted from the sequence Y, and a compressing part **17** performs entropy coding on the least significant (23−n) bits of a mantissa M of the difference signal Z, which is determined by the number of bits n following the most significant “1” in each sample in the sequence Y, and outputs a code sequence Cb. If the average amplitude is not greater than the predetermined value, the sequence X is directly losslessly coded by a compressing part **121.** A code Cd, which indicates which coding is selected, is output.

Claims(21)

1. A lossless coding method for a signal in a floating-point format, comprising:

(a) a step of converting a first signal sample sequence in the floating-point format into a second signal sample sequence in an integer format;

(b) a step of losslessly compressing said second signal sample sequence in the integer format to produce a first code sequence;

(c) a step of producing a difference signal in the floating-point format that corresponds to the difference between said second signal sample sequence in the integer format and said first signal sample sequence in the floating-point format;

(d) a step of determining which coding processing is of higher compression efficiency, a first coding processing that codes a range of bits capable of assuming non-zero in a sample of said difference signal in the floating-point format, the range of bits being determined by the bit position of the most significant “1” in the corresponding sample in said second signal sample sequence in the integer format, or a second coding processing that directly codes said difference signal; and

(e) a step of coding said difference signal by said first coding processing to produce a second code sequence if the result of said determination indicates said first coding processing, or coding said difference signal by said second coding processing to produce a second code sequence if the result of said determination indicates said second coding processing.

2. A lossless coding method for a signal in a floating-point format, comprising:

(a) a step of determining which compression is of higher compression efficiency, a direct lossless compression of a first signal sample sequence in the floating-point format on a frame basis or a dual-signal separate compression of an integer-value sample sequence and a difference signal derived from the first signal sample sequence;

(b) a step of performing the dual-signal separate compression if the result of said determination indicates the dual-signal separate compression, the step (b) comprising:

(b-1) a step of converting the first signal sample sequence into a second signal sample sequence in an integer format;

(b-2) a step of losslessly compressing said second signal sample sequence in the integer format to produce a first code sequence;

(b-3) a step of producing a difference signal in the floating-point format that corresponds to the difference between said second signal sample sequence in the integer format and said first signal sample sequence in the floating-point format;

(b-4) a step of producing a second code sequence from said difference signal in the floating-point format;

(c) a step of directly losslessly compressing said first signal sample sequence to produce a third code sequence if the result of said determination indicates the direct compression; and

(d) a step of producing an auxiliary code that indicates whether said direct lossless compression or said separate compression of two types of signals is performed.

3. A lossless coding method for a signal in a floating-point format according to claim 2 , in which, in said step (b-1), the conversion into said second signal sample sequence in the integer format is achieved by truncation of a fractional part, and

said step (b-4) comprises:

(b-4-1) a step of determining which coding processing is of higher compression efficiency, a first coding processing that codes a range of bits capable of assuming non-zero in a sample of said difference signal in the floating-point format, the range of bits being determined by the bit position of the most significant “1” in the corresponding sample in said second signal sample sequence in the integer format, or a second coding processing that directly codes said difference signal; and

(b-4-2) a step of coding said difference signal by said first coding processing to produce a second code sequence if the result of said determination indicates said first coding processing, or coding said difference signal by said second coding processing to produce a second code sequence if the result of said determination indicates said second coding processing.

4. A lossless coding method for a signal in a floating-point format according to claim 1 or 3 , in which, in said first coding processing, said range of bits capable of assuming non-zero are losslessly compressed to produce said second code sequence, or said range of bits capable of assuming non-zero are output as said second code sequence without being processed.

5. A lossless coding method for a signal in a floating-point format according to claim 1 or 3 , in which, in said second coding processing, a mantissa of said difference signal is losslessly compressed to produce said second code sequence, or said mantissa is output as said second code sequence without being processed.

6. A lossless coding method for a signal in a floating-point format according to claim 1 or 3 , in which the bit length b_{I }of one sample in said second signal sample sequence in the integer format is less than the number of bits b_{M }of a mantissa of said first signal sample sequence in the floating-point format, and in said first coding processing, said range of bits capable of assuming non-zero is separated into a higher-order sub-range containing a number of bits corresponding to the difference between the number of bits b_{M }of said mantissa and the number of bits b_{I }of said one sample and a lower-order sub-range containing the remaining bits, the sub-ranges are separately losslessly compressed to produce respective code sequences, and said code sequences are combined and output as said second code sequence.

7. A lossless coding method for a signal in a floating-point format according to claim 1 or 3 , in which said first coding processing is performed for each frame containing a plurality of samples and comprises a step of arranging factional parts, which are ranges of bits capable of assuming non-zero, of mantissas of difference signal samples in each frame with the most significant bits thereof or the least significant bits thereof aligned with each other, selecting a plurality of bits at least in the time-axis direction and/or the amplitude direction to produce a plurality of scanned bit sequences, and losslessly compressing said scanned bit sequences to produce said second code sequence.

8. A lossless coding method for a signal in a floating-point format according to claim 1 or 3 , in which said first coding processing is performed for each frame containing a plurality of samples and comprises a step of arranging a range of bits capable of assuming non-zero in each of difference signal samples in each frame and scanning the array of said bits in a plurality of different predetermined scanning orders to produce respective scanned bit sequences, a step of coding said scanned bit sequences produced by scanning in said respective scanning orders to produce code sequences for the respective scanning orders, and a step of determining which scanning order provides the code sequence of the least quantity of information and outputting the determined code sequence as said second code sequence along with an auxiliary code that indicates the corresponding scanning order.

9. A lossless coding method for a signal in a floating-point format according to claim 1 or 3 , in which said first coding processing comprises a step of determining, for each frame containing a plurality of samples, the average value of a sequence of integer-value samples, which correspond to the fractional parts of the difference signals in the relevant frame, performing error prediction, in the time-axis direction, about signals resulting from subtraction of the average value from the value of each integer-value sample, losslessly compression-coding the prediction errors, and coding the average value to form said second code sequence.

10. A coding program that makes a computer execute each step of a lossless coding method for a digital signal in a floating-point format according to claim 1 or 2 .

11. A computer-readable recording medium that stores a program according to claim 1 or 2 .

12. A lossless decoding method for a signal in a floating-point format, comprising:

(a) a step of determining whether an input code is a code produced by a single-signal coding or a dual-signal coding, based on an input auxiliary code; and

(b) if the result of said determination indicates the single-signal coding, a step of decoding and expanding an input code sequence as one code sequence to reproduce an original signal sample sequence in the floating-point format;

(c) if the result of said determination indicates the dual-signal coding, a step of separating the input code sequence into a first code sequence and a second code sequence;

(d) a step of decoding and expanding said first code sequence to produce a first signal sample sequence in an integer format;

(e) a step of deriving a difference signal in the floating-point format from said second code sequence;

(f) a step of converting said first signal sample sequence in the integer format into a second signal sample sequence in the floating-point format; and

(g) a step of combining said second signal sample sequence in the floating-point format and said difference signal in the floating-point format to reproduce the original signal sample sequence in the floating-point format.

13. A lossless decoding method for a signal in a floating-point format according to claim 12 , in which said step (e) comprises:

a step of determining which of a first expansion method and a second expansion method is used to decode and expand said second code sequence, based on said first signal sample sequence in the integer format or the input auxiliary code; and

a step of decoding and expanding said second code sequence into bit sequences each containing a number of bits capable of assuming non-zero, the number of bits being uniquely determined by the bit position of the most significant “1” in each sample of said first signal sample sequence, and assembling said difference signal from the bit sequences resulting from the decoding and expansion if the result of the determination indicates the first expansion method, or a step of directly decoding and expanding said second code sequence into said difference signal by the second expansion method if the result of the determination indicates the second expansion method.

14. A lossless decoding method for a signal in a floating-point format, comprising:

(a) a step of decoding and expanding a first code sequence to produce a first signal sample sequence in an integer format;

(b) a step of determining which of a first expansion method and a second expansion method is used to decode and expand a second code sequence, based on said first signal sample sequence or an auxiliary code;

(c) if the result of the determination indicates the first expansion method, a step of decoding and expanding said second code sequence into bit sequences each containing a number of bits capable of assuming non-zero, the number of bits being uniquely determined by the bit position of the most significant “1” in each sample of said first signal sample sequence, and assembling the bit sequences resulting from the decoding and expansion into a difference signal;

(d) if the result of the determination indicates the second expansion method, a step of directly producing a digital difference signal in the floating-point format from said second code sequence;

(e) a step of converting said first signal sample sequence in the integer format into a second signal sample sequence in the floating-point format; and

(f) a step of combining said second signal sample sequence in the floating-point format and said difference signal in the floating-point format to reproduce an original digital signal sequence in the floating-point format.

15. A lossless decoding method for a signal in a floating-point format according to claim 13 or 14 , in which the step of producing said difference signal by said first expansion method comprises:

a step of decoding and expanding a first part of said second code sequence into a bit sequence containing a number of bits capable of assuming non-zero, the number of bits being uniquely determined by the bit position of the most significant “1” in each sample in said first signal sample sequence in the integer format;

a step of decoding and expanding a second part of said second code sequence into a bit sequence containing a number of bits capable of assuming non-zero, the number of bits being uniquely determined by the difference between the number of bits of a sample in said first signal sample sequence in the integer format and the number of bits of a mantissa of an original signal sample sequence in the floating-point format; and

a step of combining the bit sequence resulting from the decoding and expansion of the first part of the second code sequence and the bit sequence resulting from the decoding and expansion of the second part of the second code sequence to produce said difference signal.

16. A decoding program that makes a computer execute each step of a decoding method for a digital signal in a floating-point format according to claim 12 or 14 .

17. A computer-readable recording medium that stores a program according to claim 12 or 14 .

18. A decoder for a signal in a floating-point format, comprising:

a first expanding part that decodes and expands an input first code sequence to produce a first signal sample sequence in an integer format;

a second expanding part that decodes and expands an input second code sequence in a difference signal into a bit sequence containing a number of bits capable of assuming non-zero, the number of bits being determined by the bit position of the most significant “1” in the corresponding sample in said first signal sample sequence in the integer format, and assembles the bits sequences resulting from the decoding and expansion into a difference signal in the floating-point format;

a third expanding part that directly decodes and expands said input second code sequence into said difference signal in the floating-point format;

a selecting part that selects one of said second expanding part and said third expanding part based on said first signal sample sequence or an input auxiliary code; and

a combining part that combines said first signal sample sequence in the floating-point format and said difference signal in the floating-point format received from the selected one of said second expanding part and the third expanding part to produce a third signal sample sequence in the floating-point format.

19. A lossless coder for a signal in a floating-point format, comprising:

first coding means of separating an input first signal sample sequence in the floating-point format into an integer value and an error and losslessly compression-coding the integer value and the error;

second coding means of directly losslessly compression-coding said first signal sample sequence in the floating-point format; and

analyzing and selecting means of determining or estimating which of said first coding means and said second coding means provides higher compression efficiency, selecting the coding means determined or estimated to provide higher compression efficiency, and producing an auxiliary code carrying information about the selection result.

20. A lossless coder for a signal in a floating-point format, comprising:

a first coding part that converts an input first signal sample sequence in the floating-point format into a second signal sample sequence in an integer format and losslessly compresses the second signal sample sequence;

a difference producing part that produces a difference signal in the floating-point format that corresponds to the difference between said second signal sample sequence in the integer format and said first signal sample in the floating-point format;

a second coding part that directly codes said difference signal in the floating-point format;

a third coding part that codes a range of bits capable of assuming non-zero in each sample of said difference signal, the range of bits being determined by the bit position of the most significant “1” in the corresponding sample in said signal sample sequence in the integer format; and

means of determining or estimating which of said second coding part and said third coding part provides higher compression efficiency, and selecting and outputting the code from the coding part determined or estimated to provide higher compression efficiency.

21. A decoder for a signal in a floating-point format, comprising:

first decoding means of directly losslessly decoding and expanding an input code sequence into a signal sample sequence in the floating-point format;

second decoding means of separating the input code sequence into an integer value and an error and losslessly decoding and expanding the integer value and the error to produce a digital signal sequence in the floating-point format; and

selecting means of selecting one of said first decoding means and said second decoding means based on an input auxiliary code.

Description

The present invention relates to a coding method for compressing a floating-point signal sample sequence carrying sound, music or image information into a code of smaller information quantity without distortion, a corresponding decoding method, a coding apparatus therefor, a decoding apparatus therefor, and programs therefor and a recording medium therefor.

As a method for compressing sound or image information, there is known a lossless coding method that involves no distortion.

Highly compressive lossless data compression can be achieved by combining a highly compressive lossy coding and a lossless compression of the difference between a reproduced signal of the lossy coding and the original signal. Such a combined compression method has been proposed in the patent literature 1. The method, which is described in detail in the patent literature, will be described briefly below.

In a coder, a frame forming part successively separates digital input signals (referred to also as an input signal sample sequence) into frames, each of which is composed of 1024 input signal samples, for example, and the digital signals are lossily compression-coded on the frame basis. This coding can be based on any format that is suitable for the input digital signal and can reproduce the original digital input signal with a certain fidelity by decoding.

For example, if the digital input signal is a sound signal, a speech coding recommended according to ITU-recommendation G.729 can be used. If the digital input signal is a music signal, a transform-domain weighted interleaved vector quantization (Twin VQ) coding used in MPEG-4 can be used. The code resulting from the lossy compression coding is locally decoded, and a difference signal that represents the difference between the locally decoded signal and the original digital signal is produced. Actually, however, there is no need of local decoding and, instead, it would suffice to obtain the difference between the original digital signal and a quantized signal resulting during the lossy compression coding. The amplitude of the difference signal is typically much smaller than that of the original digital signal. Thus, the quantity of information can be reduced by the lossless compression coding of the difference signal, compared with the lossless compression coding of the original digital signal.

To enhance the efficiency of the lossless compression coding, for a sequence of all the samples of the difference signal in a frame each expressed by sign and magnitude notation (a binary number of sign and magnitude), bits of each position, that is, bits of each of the MSB, the second MSB, . . . , and the LSB are linked along the sample sequence (that is, the time series) to form a respective bit sequence. In other words, the bit arrangement is transformed. For convenience, the bit sequence composed of linked 1024 bits at the equal bit position is referred to as a “coordinate bit sequence”. On the other hand, a one-word bit sequence representing the amplitude value of each sample including its sign is referred to as an “amplitude bit sequence”, for convenience. The difference signal has a small amplitude, and therefore, the most significant bit is, or the most significant bit and the following plural bits are, often all “0”. Thus, by representing the coordinate bit sequence formed by linking the bits at such a bit position, the efficiency of the lossless compression coding of the difference signal can be enhanced.

The coordinate bit sequence is losslessly compression-coded. As the lossless compression coding, an entropy coding, such as Huffman coding and arithmetic coding, can be used which takes advantage of the occurrence or frequent occurrence of a sequence in which the same sign (1 or 0) successively appears.

When decoding, the code resulting from the lossless compression coding is decoded, and the inverse transformation of bit arrangement is performed on the decoded signal. That is, the coordinate bit sequences are transformed into the amplitude bit sequences for each frame, and the resulting difference signals are reproduced sequentially. In addition, the code resulting from the lossy compression coding is decoded, the decoded signal and the reproduced difference signal are summed together, and then, the sum signals for each frame are linked together sequentially, thereby reproducing the original digital signal sequence.

Besides, there are known a variety of lossless coding methods for audio or visual information that permit no distortion. For example, a lossless coding method for music information is disclosed in the non-patent literature 1. Any conventional methods are such one that performs compression coding on a PCM signal directly derived from a signal waveform.

However, in music recording studios, a waveform is sometimes recorded and stored in the floating-point format. Any value in the floating-point format is separated into a sign, an exponent and a mantissa. For example, in the IEEE-754 standard floating-point format shown in

(−1)^{S}×1*.M*×2^{E−E} ^{ 0 } (1)

According to the IEEE-754 standard, E_{0 }is defined as E_{0}=2^{7}−1=127, so that the “E−E_{0}” in the expression (1) can be any value falling within the range:

−127*≦E−E* _{0}≦128.

Here, it is defined that all the bits are set at “0” when E−E_{0}=−127, and all the bits are set at “1” when E−E_{0}=128. The term “E−E_{0}=n” represents a value obtained by subtracting 1 from the number of digits (number of bits) of the integer part of the value represented by the formula (1), that is, the number of bits following the most significant “1”.

In the case where sound, music or image information is represented by a digital signal sequence in the floating-point format, the bit sequence composed of “0”s and “1”s is likely to be random because of the characteristics of the floating-point format. Thus, even if the bit arrangement transformation described above is performed, the entropy compression coding or the like cannot be expected to provide a high compression ratio. Furthermore, the sample sequence in the floating-point format significantly differs from the original analog waveform, so that there is no redundancy resulting from correlation between samples. Therefore, even if the lossless predictive coding method disclosed in the non-patent literature 1 described above is applied, a higher compression ratio cannot be expected.

Patent literature 1: JP Application Kokai Publication No. 2001-44847

Non-patent literature 1: “Lossless Compression of Digital Audio” by Mat Hans, Ronald W. Schafer et al., IEEE SIGNAL PROCESSING MAGAZINE, JULY 2001, pp. 21–32

An object of the present invention is to provide a lossless coding method and a lossless decoding method for a signal sample sequence in a floating-point format, apparatus therefor, programs therefor, and a recording medium therefor.

A lossless coding method for a signal in a floating-point format according to a first aspect of the present invention comprises:

(a) a step of converting a first signal sample sequence in the floating-point format into a second signal sample sequence in an integer format;

(b) a step of losslessly compressing the second signal sample sequence in the integer format to produce a first code sequence;

(c) a step of producing a difference signal in the floating-point format that corresponds to the difference between the second signal sample sequence in the integer format and the first signal sample sequence in the floating-point format;

(d) a step of determining which coding processing is of higher compression efficiency, a first coding processing that codes a range of bits capable of assuming non-zero in a sample of the difference signal in the floating-point format, the range of bits being determined by the bit position of the most significant “1” in the corresponding sample in the second signal sample sequence in the integer format, or a second coding processing that directly codes the difference signal; and

(e) a step of coding the difference signal by the first coding processing to produce a second code sequence if the result of the determination indicates the first coding processing, or coding the difference signal by the second coding processing to produce a second code sequence if the result of the determination indicates the second coding processing.

A lossless coding method for a signal in a floating-point format according to a second aspect of the present invention comprises:

(a) a step of determining which compression is of higher compression efficiency, a direct compression of a first signal sample sequence in the floating-point format on a frame basis or a dual-signal separate compression of an integer-value sample sequence and a difference signal derived from the first signal sample sequence;

(b) a step of performing the dual-signal separate compression if the result of the determination indicates the dual-signal separate compression, the step (b) comprising:

(b-1) a step of converting the first signal sample sequence into a second signal sample sequence in an integer format;

(b-2) a step of losslessly compressing the second signal sample sequence in the integer format to produce a first code sequence;

(b-3) a step of producing a difference signal in the floating-point format that corresponds to the difference between the second signal sample sequence in the integer format and the first signal sample sequence in the floating-point format; and

(b-4) a step of producing a second code sequence from the difference signal in the floating-point format; or

(c) a step of directly losslessly compressing the first signal sample sequence to produce a third code sequence if the result of the determination indicates the direct compression; and

(d) a step of producing an auxiliary code that indicates whether the direct lossless compression or the separate compression of two types of signals is performed.

According to the first aspect of the present invention, if the average amplitude of the input signal sample sequence X in the floating-point format is greater than a predetermined value, the signal sample sequence X is directly losslessly coded, and if the average amplitude is equal to or smaller than the predetermined value, the input signal sample sequence X is separated into an integer-value sample sequence and a difference signal, and the integer-value sample sequence and the difference signal are coded separately. Thus, the coding efficiency can be improved.

According to the second aspect of the present invention, in the case where the input signal sample sequence X in the floating-point format is separated into the integer-value sample sequence Y and the difference signal Z, which are separately coded, a case in which only the bits capable of assuming non-zero in the difference signal are coded and a case in which the difference signal is directly coded are compared in terms of compression efficiency, and one of the codes which is of higher compression efficiency is adopted. Thus, the coding efficiency can be further improved.

**12** shown in

**12** shown in

**17** that performs the coding shown in

**23** in a decoder that can decode the coding shown in

**17** in a coder according to a sixth embodiment of the present invention;

**23** in a decoder according to the sixth embodiment of the present invention;

**17** in a coder according to a seventh embodiment of the present invention;

**23** in a decoder according to the seventh embodiment of the present invention;

**17** shown in

**17** and **131** in the coder shown in

**23** and **231**, selecting part **233** and auxiliary decoding part **234** in the decoder shown in

While the present invention can be applied also to an image signal, a sound signal or the like, the following description will be made mainly with regard to embodiments in which the present invention is applied to a music signal. In addition, while various kinds of floating-point formats can be used, the following description will be made on the assumption that the 32-bit floating-point format according to the IEEE-754 standard is used.

**100**, a single-signal coding part **120**, a dual-signal coding part **110**, an auxiliary coding part **103** and an output part **104**. The analyzing and selecting part **100** comprises an efficiency estimating part **101** and a selecting part **102**. The dual-signal coding part **110** comprises an integer formatting part **12**, compressing parts **13** and **17**, and a difference producing part **14**.

For example, a signal source **10** outputs a music signal sample sequence in the form of a sequence of input digital signal samples X in the 32-bit floating-point format. The input signal sample sequence X is produced by performing a processing, such as transformation, amplitude adjustment, effect addition and mixing, on a raw signal recorded in the 24-bit integer format and converting the resulting signal, which has a fractional part as a result of the processing, into the floating-point format, or produced by converting a raw signal recorded in the 24-bit integer format into the 32-bit floating-point format and performing such processing as described above on the resulting signal. The input signal sample sequence X is divided by a frame forming part **11** into frames, each of which contains a certain number of samples, for example, 1024 samples. The input signal sample sequence may be divided into super-frames each containing 16 to 32 frames, or each of the resulting frames may be subdivided into halves or into quarters. It is essential only that the input signal sample sequence is divided into a group of a certain number of samples, and such a group of samples will be generically referred to as a frame.

The input signal sample sequence X is analyzed by the analyzing and selecting part **100** and then passed to any one of the dual-signal coding part **110**, which losslessly compression-codes an integer value and an error thereof separately, and the single-signal coding part **120**, which directly losslessly compression-codes the input signal sample sequence X. In the analyzing and selecting part **100**, for example, the input signal sample sequence X is input to the efficiency estimating part **101**, where an averaging part **101** *a *calculates the average amplitude for the relevant frame, and a determination part **101** *b *determines whether or not the average amplitude is equal to or greater than a predetermined value, and the selecting part **102** is controlled based on the result of the determination. If the average amplitude is determined as being equal to or greater than the predetermined value, the input signal sample sequence X of the frame is supplied to the dual-signal coding part **110**, or if the average amplitude is determined as being smaller than the predetermined value, the input signal sample sequence X of the frame is supplied to the single-signal coding part **120**.

For example, if the integer part of each sample in the input signal sample sequence X is represented by a 16-bit binary number, the predetermined value used as a criterion for the determination in the determination part **101** *b *is a 8-bit value. The predetermined value varies with the integer value described above, and is selected based on the characteristics of the input signal sample sequence so that the entire compression efficiency is improved.

In the dual-signal coding part **110**, the signal sample sequence X in the floating-point format is input to the integer formatting part **12** and converted into a signal sample sequence Y in the integer format on a sample basis. As described above, in the example described above, the digital signal sample X is often derived from a raw signal in the 24-bit integer format through a processing, such as transformation, amplitude adjustment and effect addition, so that the amplitude of the signal typically does, not vary significantly. Thus, the integer formatting part **12** can convert the signal sample sequence X into the integer format simply by rounding, truncating or rounding up the fractional part.

As an example, truncation will be described with reference to

To represent a waveform in the floating-point format, normalization may be performed so that an integer value of 32768 (=2^{15}) is normalized to 1.0, or an integer value of 2^{23 }is normalized to 1.0. However, these cases differ from the case described below only in the value of the exponent but don't have an essential difference.

If the exponent value E is 147, for example, the bit adjacent the decimal point on its higher order side is the 20th (147−120=20) bit from the MSB in the mantissa. As shown in _{I}) in the 23-bit mantissa M correspond to a part that follows the most significant “1” of the integer part in the sign and magnitude binary notation, and the remaining three least significant bits (M_{2}) in the mantissa M correspond to the fractional part below the decimal point. In the following, the part M_{I }will be referred to as an integer part of the mantissa M. Therefore, as shown in _{I}) of the mantissa is positioned at the least significant bit in the entire 23-bit mantissa, overflow of the three bits (M_{2}) below the decimal point occurs, and thus, the three bits are truncated. Then, the least significant bit in the most significant three bits having been made unoccupied by such shifting (“000” in this example) (that is, the 21st bit from the least significant bit) is set at 1, which corresponds to “1” of “1.M” in the expression (1), thereby providing a truncated integer value (see

Furthermore, the resulting integer value is converted into a two's complement notation. That is, the sign bit S of each signal sample sequence X in the floating-point format is used as it is as the most significant bit, and as for the other 23 bits, if the sign S is “0” (positive), the 23 bits shown in

The above description has been made generally on the assumption that the mantissa M contains 0 or more bits corresponding to the fractional part, and a 32-bit digital signal sample in the floating-point format is converted into a 24-bit digital signal sample in the integer format. However, for example, in the case where a plurality of raw signals in the 24-bit integer format are mixed for processing, one sample may have an amplitude value that is significantly greater than the maximum value that can be represented by 24 bits. In such a case where E≧150, as an exceptional processing, the value of the exponent E is fixed to 150 (E=127+23=150) in the example described above, an exception signal y is provided to the difference producing part **14**, and the most significant 23 bits of the “1M”, which is the mantissa M with “1” added at the top thereof, is converted into the two's complement notation to produce a signal sample sequence Y in the 24-bit integer format.

The integer formatting part **12** performs a processing shown in **1**). If the integer part is constituted by less than 23 bits, the digit calculating part **12**A determines the number of digits n (n=E−E_{0}) of the integer part of the mantissa and outputs the result (S**2**). Then, the mantissa M is shifted by (23−n) bits toward the least significant position to make overflow of the fractional part occur, and “1” is placed at the (n+1)th bit position viewed from the least significant bit of the resulting integer part composed of n bits (S**3**). Then, “0”s are placed at the remaining bit positions preceding the (n+1)th bit so that the whole mantissa is composed of 23 bits, and based on the sign bit S, the resulting 23 bits are converted into the 24-bit integer format represented in the two's complement notation, thereby producing a signal sample Y in the integer format (S**4**). Specifically, the sign bit S is used as the most significant bit as it is, and as for the remaining 23 bits, if the sign S is “0” (positive), the bits, from the least significant bit to the 23rd bit, of the integer part resulting from the shifting and the addition of “1” to the (n+1)th bit from the least significant bit in step S**3** are used as they are, and if the sign S is “1” (negative), the bits, from the least significant bit to the 23rd bit, of the integer part resulting from the shift and the addition of “1” are inverted, and “1” is added to the least significant bit of the resulting 23 bits to provide a value in the complement notation. If the exponent E is equal to or greater than 150 in step S**1**, the exponent E is set at 150 in step S**5**, and then, a signal sample Y in the integer format is produced from the most significant 23 bits of the “1M” in step S**4**.

In the above description, the mantissa M is shifted by (23−n) bits to provide a 24-bit absolute value sample in the sign and magnitude binary notation. Alternatively, most significant n bits (n=E−E_{0}) of the mantissa M shown as M_{I }in **23**, and then the sign bit S be added at the top of the 23 bits to make the number of bits **24**.

The integer-value signal sample sequence Y produced by the integer formatting part **12** as described above is, as an integer value, compression-coded in the compressing part **13** by an efficient lossless compression method using correlation of a waveform or the like, and then output as a code sequence Ca. For example, as disclosed in the non-patent literature 1 described above, the lossless compression in the compressing part **13** may be achieved by determining the difference between the value of each sample and a predicted value (integer value), transforming the bit arrangement of a sequence of such differences as described in the part “BACKGROUND ART”, and then performing the entropy coding on the resulting sequence, that is, the coordinate bit sequence. That is, the integer-value signal sample sequence Y approximates to the original analog signal waveform from which the input signal sample sequence X from the signal source **11** has been derived. Therefore, the input signal sample sequence X can be efficiently losslessly compression-coded by eliminating, by prediction or transformation, the redundancy due to the correlation among the signal samples.

The difference producing part **14** produces a difference signal (error) Z, which indicates the difference between the integer-value signal sample Y and the corresponding input signal sample X in the floating-point format from the signal source **11**. In this example, the floating-point formatting part **15** converts the integer-value signal sample Y again into a signal sample Y′ in the floating-point format, and the subtraction part **16** subtracts the resulting signal sample Y′ in the floating-point format from the original input signal sample X in the floating-point format, thereby producing the difference digital signal Z in the floating-point format.

In the case where the digital signal sample in the integer format is composed of 24 bits, the floating-point formatting part **15** can convert the digital signal into an input signal in the 32-bit floating-point format without any ambiguity or exception. As described above, the exponent E of the original signal sample sequence X in the floating-point format is likely to be equal to or less than 149, and in such a case, a difference signal sample Z indicating the difference between a sample in the signal sample sequence Y′ and a sample in the original input signal sample sequence X in the floating-point format equals to the value of the fractional part of the sample in the original input signal sample sequence X.

The above description concerning the difference producing part **14** is intended only to help understanding of the present invention, and actually, there is no need of converting the integer-value signal sample sequence Y into the floating-point format. The input sample sequence X and the number of digits n of the integer part of the mantissa thereof are input to the difference producing part **14**, and if no exception signal y is provided, the sign S, the exponent E and the (23−n) least significant bits of the mantissa M of each sample in the input sample sequence X in the floating-point format are used as the sign S, the exponent E and the (23−n) least significant bits of the mantissa of each difference signal sample Z, respectively, and the remaining most significant n bits of the mantissa are all set at “0”. The number of digits, n (n=E−E_{0}), of the integer part of the mantissa equals to the number of digits following the most significant “1” of the corresponding sample in the integer-value signal sample sequence Y. Since the bit length of the integer-value signal sample is fixed, the number of digits of the integer part of the mantissa is also determined when the number of digits preceding the most significant “1” is determined. That is, if the position of the most significant “1” in an integer-value signal sample is found, the number of digits of the integer part of the mantissa of the corresponding floating-point signal sample can also be found. If an exception signal y is provided, the sign of the difference signal sample Z is constituted by the sign S of the input sample, the exponent of the difference signal sample Z is set at E−150, and the mantissa of the difference signal sample Z is constituted by the difference between the mantissa of the input sample sequence X and the signal sample sequence Y′.

The compressing part **17** performs a lossless compression coding on the difference signal Z for each sample combination of an integer-value signal sample sequence Y and an input sample sequence X in the floating-point format, that is, the difference signal Z in the floating-point format from the difference producing part **14** and outputs a code sequence Cb. For example, the compressing part **17** performs the lossless compression coding only on the digit(s) in the difference signal Z in the floating-point format that can assume a value other than “0” for each sample, using the number of digits n of the integer part of the mantissa that is output from the integer formatting part **12**. Each sample in the input sample sequence X in the floating-point format is expressed as follows:

*X=*1*.M×*2^{n} *=X* _{W} *.X* _{F }

Provided that X_{W}=1M_{I }(M_{I }with “1” added at the top thereof) indicates the integer part in the binary notation, and X_{F}=M_{F }indicates the fractional part, the bits corresponding to the integer part M_{I }in the mantissa M and the bits corresponding to the fractional part (decimal value) M_{F }are related to each other as shown in _{D }in the sign and magnitude binary notation is shown for convenience.

When the absolute value X_{W }of the integer part is 0, the expression (1) can be classified into two cases: a case of E−E_{0}=−127; and a case of −126≦E−E_{0}≦−1. The former case is a special case of E=0, in which all the 23 bits “x_{23 }. . . x_{1}” of the mantissa M constitute the fractional part M_{F}, and the 23 bits of the mantissa of the sample in the input signal sample sequence X are all “0” by definition. In the case of −126≦E−E_{0}≦−1, the mantissa constitutes the fractional part, but not all the bits thereof are “0”. In any case, in the case where E−E_{0 }assumes a negative value, all the bits of the sample in the integer-value signal sample sequence Y are “0”, and the difference signal Z is the same as the sample in the input sample sequence X, so that all the 32 bits of the sample in the signal sample sequence X in the floating-point format are to be coded.

When the absolute value X_{W }of the integer value is 1, the bit “1” at the position one bit higher than the 23rd bit of the mantissa M (referred to as a 24th bit, hereinafter) represents the integer value X_{W}, and the 23 bits “x_{23 }. . . x_{1}” of the mantissa M represent the fractional part M_{F}.

When the absolute value X_{W }of the integer value is 2 to 3, the two bits “1x_{23}” composed of the 24th bit “1” and the 23rd bit x_{23 }represent the integer value X_{W}, in which the bit x_{23 }represents the integer part M_{I }in the mantissa M, and the remaining 22 bits “x_{22 }. . . x_{1}” represent the fractional part M_{F}. When the absolute value X_{W }is 4 to 7, the three bits “1x_{23}x_{22}” composed of the 24th bit “1”, the 23rd bit x_{23 }and the 22nd bit x_{22 }represent the integer value X_{W}, in which the bits “x_{23}x_{22}” represent the integer part M_{I }in the mantissa M, and the remaining 21 bits “x_{21 }. . . x_{1}” represent the fractional part M_{F}.

When the absolute value X_{W }is 2^{k−1 }to 2^{k}−1, the k bits composed of the 24th bit “1” and the bits “x_{23 }. . . x_{23-(k-2)}” represent the integer value X_{W}, in which the bits “x_{23}x_{22 }. . . x_{23-(k-2)}” represent the integer part M_{I }in the mantissa M, and the remaining (23−k) bits “x_{23-(k-1) }. . . x_{1}” represent the fractional part M_{F}.

Since the difference signal Z is produced by subtracting the sample sequence Y′ in the 32-bit floating-point format converted from the integer-value signal sample sequence Y from the input sample sequence X in the 32-bit floating-point format, the difference signal Z is as shown in _{W }is 0, which corresponds to the above-described case where E−E_{0 }assumes a negative value, the sign S, the exponent E and the mantissa M of the difference signal Z are all the same as those of the input sample X.

In the case where the absolute value X_{W }of the integer value is 1, the sign S and the exponent E of the sample in the sample sequence Y′ are the same as those of the corresponding sample in the sample sequence X, and the mantissa M of the sample is 0. In this case, subtraction of exponents is also performed. Thus, the sign S and the exponent E of the difference signal Z are always the same as those of the corresponding sample in the sample sequence X, that is, 0 and 0, respectively, and the mantissa M of the difference signal Z is the same as the mantissa M of the sample in the sample sequence X.

In the case where the integer value X_{W }is 2 to 3, the sign S and the exponent E of the sample Y′ are the same as those of the sample X, and as for the mantissa M of the sample Y′, the 23rd bit x_{23 }thereof is the same as that of the mantissa M of the sample X, and the bits constituting the fractional part “x_{22 }. . . x_{1}” thereof are all “0”. Therefore, the sign S and the exponent E of the difference signal are always 0, the most significant bit (the 23rd bit) in the mantissa M is always 0, and the remaining bits constituting the fractional part “x_{22 }. . . x_{1}” are the same as the respective corresponding bits of the sample X and vary depending on the sample X.

Similarly, in the case where the integer value X_{W }of the sample X is 2^{k−1 }to 2^{k}−1, the sign S and the exponent E of the difference signal Z are always 0, and as for the mantissa M thereof, the most significant (k−1) bits “x_{23 }. . . x_{23-(k-2)}” corresponding to the integer part of the mantissa M of the sample X are always 0, and the remaining 23−(k−1) bits “x_{23-(k-1) }. . . x_{1}” are the same as the respective corresponding bits of the mantissa M of the sample X and vary depending on the sample X. In this way, the number of digits h (h=23−n) of the difference signal Z that can be non-zero is determined depending on the range of the integer value of the signal sample sequence Y in the integer format, that is, the number of digits, n+1, of the integer value.

Because of such a relationship, the compressing part **17** performs a lossless compression coding only on the least significant h (h=23−n) bits that can be non-zero in the difference signal Z in the floating-point format, using the number of digits n of the integer part of the mantissa input from the integer formatting part **12**. That is, if the number of digits n of the integer part of the mantissa is 0, all the bits of the exponent E and the mantissa M are losslessly compression-coded, and if the number of digits n is not 0 (n≠0), only the least significant h (h=23−n) bits of the mantissa M are losslessly compression-coded. From the above description, it will be understood that the difference signal Z can be produced by inputting the input signal sample X and the number of digits n of the integer part of the mantissa to the difference producing part **14**.

As described above, when the value of E−E_{0 }in the expression (1) of the sample in the input signal sample sequence X is from −127 to 0, the difference signal Z is the same as the input signal sample X. Therefore, the compressing part **17** codes the 32 bits of the input signal sample X as the difference signal Z. When the value of E−E_{0 }is 1 to 22, in the case where the difference signal sample sequence Z (=z_{1}, z_{2}, z_{3}, . . . ) is as shown in _{D }constituted by the bits that can be non-zero is coded, except the case where n≠0. The coding of the bits that can be non-zero can be achieved by entropy coding, such as Huffman coding, of h bits that can be non-zero, or by universal coding of the whole of the bits. That is, only the part shaded in _{D }in each frame are coded as a series of bit sequences, the series of bit sequences decoded on the decoder side can be separated again based on the number of bits h capable of assuming non-zero in each sample.

The sign S is contained in the code Ca as a result of coding of the signal sample sequence Y in the integer format, and therefore, there is no need of coding and transmitting the sign S when coding the difference signal Z. In addition, since the n (=E−E_{0}) can be determined from the number of bits n following the most significant “1” except the sign bit of each sample in the integer-value signal sample sequence Y on the receiver side, there is no need of transmitting the number of bits n.

The difference signal Z sometimes has a large number of digits following the decimal point, and in such a case, the coding efficiency can be reduced. Thus, in the embodiment shown in

As described above, in the exceptional case where the mantissa E of the signal sample sequence X in the floating-point format holds E≧150, the integer formatting part **12** provides an exception signal y, which indicates that the exponent E is fixed to 150, to the difference producing part **14**, and the difference producing part **14** produces a difference signal Z in the floating-point format whose exponent is constituted by the difference (E−150) between the fixed exponent value 150 and the exponent E of the sample in the signal sample sequence X and whose mantissa is constituted by the difference in mantissa M. The compressing part **17** losslessly compression-codes the difference signal Z and outputs the resulting code sequence Cb.

As shown in **120**, the input signal sample sequence X in the floating-point format is directly losslessly compression-coded in a compressing part **121**, and output therefrom as a code sequence Cc. The lossless compression coding can be achieved by 8-bit universal coding, for example. The auxiliary coding part **103** in the analyzing and selecting part **100** codes auxiliary information as to which of the dual-signal coding part **110** and the single-signal coding part **120** has been selected and outputs the information as an auxiliary code Cd. The code sequences Ca and Cb and the auxiliary code Cd, or the code sequence Cc and the auxiliary code Cd, are output from the output part **104** to be transmitted or recorded in a recording medium.

Code sequences are input at an input terminal **201**, and an auxiliary code extracting part **202** extracts the auxiliary code Cd from each frame of code sequences. A separating part **203** is controlled based on the auxiliary code Cd. If the auxiliary code Cd indicates the dual-signal coding, the separating part **203** separates the code sequences Ca and Cb received via the input terminal **201** and passes the code sequences to a dual-signal decoding part **210**, or if the auxiliary code Cd indicates the single-signal coding, the separating part **203** separates the code sequence Cc from the code sequence received via the input terminal and passes the code sequence Cc to a single-signal decoding part **220**.

The input code sequence Ca input to the dual-signal decoding part **210** is losslessly expansion-decoded in an expansion part **21**. This lossless expansion decoding corresponds to the lossless compression coding performed in the compressing part **13** shown in **21**A in the expansion part **21** determines the number of digits n following the most significant “1” in each expansion-decoded sample as the number of digits of the integer part of the mantissa of the corresponding digital signal in the floating-point format and outputs the number of digits n. The number of digits n is input to an expansion part **23**.

The expansion part **23** losslessly expansion decodes the input code sequence Cb. This lossless expansion decoding corresponds to the lossless compression coding performed in the compressing part **17** shown in **23**, an assembly part **23**A assembles a sequence of expansion-decoded bits into a difference (error) digital signal in the 32-bit floating-point format for each sample based on the number of digits n, and outputs the resulting digital signal as the difference signal Z.

Based on the number of digits n for the first sample in the frame provided from the expansion part **21**, if the number of bits n for the first sample does not hold n=0, the assembly part **23**A extracts the least significant h (h=23−n) bits from the 23-bit mantissa M of the bit sequence expansion-decoded by the expansion part **23** and assembles the extracted h bits into a difference signal z_{1 }in the 32-bit floating-point format, which is constituted by the shaded part of the first sample z_{1 }shown in _{I }of the mantissa M corresponding to the integer part shown in _{2}, z_{3}, . . . shown in

The signal sample sequence Y in the 24-bit integer format from the expansion part **21** is converted into a signal sample sequence Y′ in the 32-bit floating-point format by the floating-point formatting part **22**. A combining part **24** combines the samples in the signal sample sequence Y′ with the respective difference signals Z in the floating-point format from the expansion part **23** to reproduce the signal sample sequence X in the floating-point format. The sign of each sample in the reproduced signal sample sequence X is the same as that of the corresponding sample in the signal sample sequence Y′, that is, the same as the sign of the corresponding sample in the decoded signal sample sequence Y in the integer format. Except for the case where n=0, the exponent E of each sample in the sample sequence X is the same as that of the corresponding sample in the signal sample sequence Y′, the most significant n (n=23−h) bits of the mantissa M of each sample in the sample sequence X are the same as the most significant n bits of the mantissa M of the corresponding sample in the signal sample sequence Y′, and the remaining, least significant h (h=23−n) bits of each sample in the sample sequence X are the same as the least significant h bits of the difference signal Z.

The code sequence Cc input to the single-signal decoding part **220** is directly decoded into the original signal sample sequence X in the floating-point format through a lossless expansion-decoding process corresponding to the compression process performed in the compressing part **121** shown in

The signal sample sequences X in the floating-point format decoded by the dual-signal decoding part **210** and the single-signal decoding part **220** are sequentially combined into a signal sample sequence in the floating-point format by a frame combining part **205**.

**100** shown in **110** and the single-signal coding part **120** code the input signal sample sequence X and input the code sequences Ca and Cb and the code sequence Cc, respectively, to an analyzing and selecting part **100**′, in which the compression efficiency for each frame is calculated by an efficiency calculating part **100** *c*. For example, the total code length for the relevant frame is calculated for each of the set of code sequences Ca and Cb and the code sequence Cc, and a comparison and determination part **100** *d *compares the total code lengths with each other, and determines the code sequence(s) of the shorter total code length as the code sequence(s) compressed with higher efficiency. The determination result is passed to the auxiliary coding part **103**, and the auxiliary coding part produces a code Cd. In addition, the output part is controlled based on the comparison result and selectively outputs a code set (Ca, Cb, Cd) or a code set (Cc, Cd). The remaining configuration is the same as that shown in

In the coder shown in **17** directly losslessly compression-codes the difference signal Z in the floating-point format, the expansion part **23** in the decoder shown in **12** may perform integer-formatting by not only truncation but also rounding and round-up.

There are sometimes cases where an original 24-bit or 16-bit integer-value PCM signal is converted into a floating-point number for the purpose of convenience to form a sequence to be coded instead of using an ordinary floating-point number signal to produce a sequence to be coded. In the case of such a special floating-point sequence, the bits that can be non-zero described above with reference to the first embodiment disappear. Thus, the sign S, the exponent E and the mantissa M of the difference (error) signal Z in the floating-point format from the difference producing part **14** are all zero, and there is no need of transmitting the difference signal Z. In this case, for example, two bits are used as auxiliary information, and only needed is to inform the decoder that, if the two bits are “00”, it means that the original digital signal is a 16-bit integer-value signal sample, if the two bits are “01”, it means that the original digital signal is a 24-bit integer-value signal sample, and if the two bits are “10”, it means that the original digital signal is a signal sample other than those described above. To this end, the 2-bit auxiliary information is added at the top of the code sequence Ca output from the compressing part **17**. Efficient compression can be achieved if only the 2-bit auxiliary code is output as the code sequence Cb in the former two cases, or the auxiliary code is output together with the code sequence Cb in the latter case.

According to a second embodiment of the present invention, a signal sample sequence X in the floating-point format is always converted into a signal sample sequence Y in the integer format. However, the difference signal Z in the floating-point format is coded by one of two lossless compression coding processes which is expected to provide a higher compression efficiency. Specifically, one is a process of losslessly coding only the bits capable of assuming non-zero in the difference signal in the floating-point format according to the number of digits of each sample in the signal sample sequence Y in the integer format as described with reference to the first embodiment. The other is a process of directly losslessly compression-coding the difference signal in the floating-point format.

**17** that uses the number of digits n following the most significant “1” in the integer-value sample and a compressing part **131** that directly losslessly compression-codes the difference signal Z in the floating-point format. The number of digits n of the integer part of the mantissa M (the number of digits following the most significant “1” in the integer-value sample) is input to an analysis part **132** from an integer formatting part **12**. A determination part **132** *a *compares the number of digits n with a reference value, and a selecting part **133** is controlled based on the comparison result. The most significant n bits of the mantissa M of the difference signal Z are “0”, and as the number of digits n increases, the compression efficiency of the compressing part **17** increases. Thus, if the number of digits n is equal to or greater than the reference value, the difference signal Z is supplied to the compressing part **17**, or if the number of digits n is smaller than the reference value, the difference signal Z is supplied to the compressing part **131**. Then, a code sequence Cb corresponding to the difference signal Z losslessly compression-coded by the compression part **17** or **131** is to be input to an output part **104**.

If the number of digits n is large, that is, if the amplitude is large, the number of bits h (=23−n) capable of assuming non-zero in each difference signal sample Z is small as described above with reference to **131** that losslessly compression-codes the difference signal Z in the floating-point format by the typical 8-bit universal coding, for example. For example, one bit of “0” is added at the top (or the bottom) of the 23-bit mantissa of each difference signal sample to forma a 24-bit bit sequence, the bit sequence is divided into 3 symbols (i.e., coding units) each containing 8 bits, and each of the 8-bit symbols is compression-coded. Alternatively, only the bits capable of assuming non-zero in the mantissa of each difference signal sample may be separated on an 8-bit basis. If the number of the bits is not divisible by 8, a required number of bits of “0” are added to the bit sequence, the resulting bit sequence is divided into 8-bit symbols, and each of the symbols is compression-coded. For example, if the number of bits capable of assuming non-zero in the mantissa of a sample is 11, five bits of “0” are added to the bit sequence, and the resulting bit sequence is divided into two 8-bit symbols.

For example, the reference value used in the determination part **132** *a *is 10, if the amplitude length of the digital signal sample sequence Y in the integer format is 16 bits. As with the predetermined value used in the determination part **101** *b *in the first embodiment, the reference value is determined based on various conditions so that high compression efficiency is achieved.

Selection between the compressing parts **17** and **131** may be performed for each sample. Alternatively, the selection may be performed for each frame in such a manner that an averaging part **132** *b *in the analysis part **132** calculates the average number of digits n following the most significant “1” for the integer-format signal samples in each frame, and the compressing part **17** is selected if the average is equal to or greater than a reference value, or the compressing part **131** is selected if the average is smaller than the reference value. In the case where selection between the compressing parts **17** and **131** is performed for each sample, for example, the compressing part **17** couples the bits capable of assuming non-zero in each of the selected samples in the input frame, and performs entropy coding of the coupled bits. Since the number of digits n can be determined from the signal sample in the integer format decoded on the decoder side, even when the selection between the compressing parts **17** and **131** is performed based on the number of digits n, there is no need of transmitting an auxiliary code that indicates which of the compressing parts **17** and **131** has been selected.

**203** is always input to an expansion part **21** and decoded and expanded into a signal sample sequence Y in the integer format. For expanding the code sequence Cb, an expansion part **231** is provided in addition to the expansion part **23**.

The number of digits n following the most significant “1” in the digital signal sample in the integer format from the expansion part **21** is input to an analysis part **232**. The analysis part **232** is configured the same as the analysis part **132** shown in **232** *a *compares the number of digits n with a reference value, and a selecting part **233** is controlled so as to supply the code sequence Cb to the expansion part **23** if the number of digits n is equal to or greater than the reference value, and to supply the code sequence Cb to the expansion part **231** if the number of digits n is smaller than the reference value. As described above with regard to the first embodiment, the expansion part **23** extracts h bits capable of assuming non-zero from the decoded and expanded bit sequence, the number h being determined by the number of digits n, and an assembly part **23** *a *assembles the extracted bits into a difference signal Z in the floating-point format.

The expansion part **231** directly decodes and expands the code sequence Cb into a difference signal Z in the floating-point format through a decoding and expanding process corresponding to the process performed in the compressing part **131** shown in **132** shown in **232** *b *in the analysis part **232** calculates the average of the numbers of digits n, which is compared with a reference value. Based on the comparison result, the selecting part **233** is controlled to supply the code sequence Cb to the expansion part **23** or **231**. The expansion part **23** decodes the code sequence Cb into a bit sequence composed of bits capable of assuming non-zero, as in the case described with reference to

As described above, in this embodiment, although the difference signal Z is in the floating-point format, all the bits of the exponent and sign of the difference signal Z in the floating-point format produced by the difference producing part **14** in the coder are “0”. Therefore, the exponent and sign are not transmitted to the decoder, and the decoder assembles the mantissa of the difference signal Z from the bits capable of assuming non-zero decoded by the expansion part **23** and then assembles the difference signal Z in the floating-point format by adding the exponent composed of “0” bits and the sign bit “0” to the mantissa.

The expansion part **231** receives the code sequence Cb as a code sequence Cb′ and performs a decoding process which is the inverse of the coding performed in the compressing part **131** shown in **24** sums any of these difference signals Z and the sample Y′ in the floating-point format corresponding to the sample Y in the integer format, thereby forming the signal sample X in the floating-point format.

**132** and the selecting part **133** shown in **17** and **131**, it is determined which of the compressing parts can achieve coding with higher compression efficiency, and the code produced by the compressing part of higher compression efficiency is output. Specifically, the difference signal Z may be losslessly compression-coded by both the compressing parts **17** and **131**, the resulting code sequences Cb and Cb′ may be input to an analyzing and selecting part **134**, an efficiency calculating part **134** *a *and a determination part **134** *b *may determine which of the compressing parts **17** and **131** codes the difference signal Z with higher compression efficiency for the relevant frame, and the code sequence of higher compression efficiency may be output as the code sequence Cb from an output part **104**. In this case, an auxiliary code Ce that indicates which of the compressing parts **17** and **131** has produced the code sequence Cb has to be output from an auxiliary coding part **103**.

A decoder suitable for the coder shown in **232** removed and an auxiliary decoding part **234** shown by a broken line provided. The auxiliary decoding part **234** decodes the auxiliary code Ce separated by the separating part **203**, and based on the decoding result, the selecting part **233** is controlled to supply the code sequence Cb to the expansion part **23** or **231**. The remaining operation is the same as that of the decoder adapted for the coder shown in

The analyzing and selecting part **134** and the auxiliary coding part **103** shown in **234** may be provided in the decoder shown in **232**, so that any of the determination methods can be used. However, corresponding determination methods have to be used in the coder and the decoder.

A third embodiment of the present invention is a combination of the first embodiment shown in **8**, **9** and **11** are denoted by the same reference numerals as in those drawings, and redundancy of the descriptions thereof will be avoided. According to this embodiment, depending on the combination of selections made by selecting parts **102** and **133**, any of code sequence sets (Ca, Cb, Cd), (Ca, Cb′, Cd) and (Cc, Cd) is selected and output.

Selection from among such three code sequence sets can be made by additionally providing the compressing part **121** shown in **134**.

A decoder according to the third embodiment is shown in

According to the first embodiment described above, there is provided a coder that converts the input signal sample X in the floating-point format into the integer format by truncating the fractional part thereof and codes only the bits capable of assuming non-zero in the difference signal Z taking advantage of the fact that all the bits of the integer part in the mantissa of the difference signal Z are “0”, thereby improving the compression efficiency. In addition, according to the second embodiment, there is provided a coder that chooses the higher one of the compression efficiency in the case where only the bits capable of assuming non-zero are coded and the compression efficiency in the case where all the bits are coded. According to a fourth embodiment, in addition to truncating the part following the decimal point, a predetermined number of bits of the integer part are also truncated to produce an integer-value signal sample Y. In the example described below, there will be shown a case where, in the conversion into a 16-bit integer value through truncation of bits following the decimal point, 8 bits are further truncated. However, the present invention can be applied to any conversion into an integer value composed of 17 bits to 23 bits.

According to the fourth embodiment, the input signal sample X in the floating-point format can be coded through the same process regardless of whether the original signal from which the input signal sample X has been derived is a floating-point signal, a 24-bit integer-value signal or a 16-bit integer-value signal. However, for convenience of explanation of the principle of operation, a case where the input signal sample X is derived from a floating-point signal, a case where the input signal sample X is derived from a 24-bit integer-value signal containing one sign bit, and a case where the input signal sample X is derived from a 16-bit integer-value signal containing one sign bit will be described in order.

With reference to _{X}” represents 23 bits of the mantissa of the input signal sample X in the floating-point format, and “P_{D}” represents the position of the decimal point that indicates the boundary between the integer part constituted by the most significant n (n=E−E_{0}) bits and the fractional part. By reducing the exponent E of the sample X to E−8, the position of the decimal point PD of the mantissa M_{X }is shifted toward the most significant bit position by 8 bits. As a result, in the example shown in _{X }constitute the integer part of the mantissa after shifting. Then, a bit “1” is added at the top of the 2 bits constituting the integer part, and the resulting sequence of bits “1xx” is used as the least significant bits of a 16-bit integer-value signal sample Y.

Each sample in the 16-bit integer-value signal sample Y is shifted toward the most significant bit position by 8 bits, that is, 8 bits of “0” are added at the bottom of the 16-bit integer-value signal sample Y, and then, the resulting signal sample sequence is converted into a signal sample sequence Y′ in the floating-point format. Thus, the exponent E of the resulting signal sample sequence Y′ in the floating-point format has been added with 8 and equals to the exponent E of the input signal sample sequence X. The mantissa My of the signal sample sequence Y′ is formed by setting, as the most significant bits, the bits of the integer part that follow the most significant “1” of the integer-value signal sample sequence Y having been shifted toward the most significant bit position by 8 bits, that is, the 2 bits “xx” in this example, and setting all the following bits at “0”. For convenience, _{Y′} of the signal sample sequence Y′ in the floating-point format resulting from conversion. The mantissa M_{Z }of the difference signal Z is the difference between the mantissa M_{X }and the mantissa M_{Y′}. The most significant (n−8) bits of the mantissa M_{Z }of the difference signal Z, or the most significant 2 bits in this example, are “00”, and the following bits (fractional part) equal to the respective corresponding bits of the mantissa M_{X}. That is, all the bits that follow the most significant bits “xx” constituting the integer part of the mantissa M_{Z }can be non-zero.

As for the difference signal Z, only the least significant 23−(n−8) bits of the mantissa M that can be non-zero have to be coded. The number of the most significant (n−8) bits of the mantissa M_{Z }that are always set at 0 is the same as the total number of bits following the most significant “1” in the integer-value signal sample sequence Y decoded on the receiving side. As shown in

_{X }of the sample sequence X differs from the mantissa M_{X }shown in _{D }are “0” as shown in _{D }toward the most significant bit position by 8 bits. Then, the resulting 16-bit integer-value signal sample sequence Y is the same as in the case shown in _{Y′} of the signal sample sequence in the floating-point format converted from the integer-value signal sample sequence Y is also the same as in the case shown in

However, the mantissa M_{Z }of the difference signal is the difference between the mantissas M_{X }and M_{Y′}, and as shown in _{Z }of the difference signal Z is “00”, the following 8 bits represents a range of bits that can be non-zero, and the bits further following the 8 bits are all “0”. The range of bits that can be non-zero is shifted toward the least significant bit position as the number of bits constituting the integer part increases, as shown in

_{X}, the 8 bits following the decimal point PD are all “0”, and the bits following the 8 bits are also all “0”. While the number of digits h following the decimal point varies with the value of n according to h=23−n, the case shown in

In _{D }toward the most significant bit position by 8 bits, and a 16-bit integer-value signal sample sequence Y is produced using the integer part “xx” of the mantissa after shifting in the same manner as shown in _{Y′} of the sample sequence in the floating-point format converted from the sample sequence Y is also the same as in the case shown in _{Z }of the difference signal Z are all “0”, the following 8 bits represents a range of bits that can be non-zero, and the bits further following the 8 bits are all “0”. In this case also, as shown in

As can be apparent from the above description, whether the input signal sample in the floating-point format is derived from a 24-bit integer-value signal, a 16-bit integer-value signal or a floating-point signal, the 16-bit integer-value signal sample Y can be produced through 8-bit truncation by shifting the decimal point toward the most significant bit position by 8 bits by reducing the exponent E of the input signal sample X to E−8, adding “1” at the top of the most significant (n−8) bits constituting the integer part of the mantissa M_{X}, or the bits “xx” in the example shown in the drawings, and adopting the resulting integer value “1xx” as the least significant bits of the 16-bit integer-value signal sample Y. In addition, the mantissa M_{Z }of the difference signal Z can result from setting the most significant (n−8) bits of the mantissa M_{X }of the input signal sample X at “0”.

The 16-bit integer-value signal sample Y thus obtained can be compression-coded in the same manner as in the embodiment shown in _{Z }that are always “0” are not coded, and the following 8 bits that can be non-zero and the remaining (23−n) bits further following the 8 bits are compression-coded separately.

In the examples described above, the 16-bit integer value is produced by shifting the decimal point in the 23-bit mantissa toward the most significant bit position by 8 bits and truncating the bits following the shifted decimal point. In general, m bits (m: any integer falling within a range 22≧m≧8) can be truncated by reducing the exponent E of the input signal sample to E-m and adopting the most significant (n-m) bits of the resulting mantissa M_{X }as the integer part. In addition, the mantissa of the difference signal Z can be coded efficiently if the most significant (n-m) bits that are always “0” are not coded, and the following m bits that can be non-zero and the remaining (23−n) bits further following the m bits are coded separately.

**14**B and **14**C, a digit adjusting part **31** reduces the exponent E of the input signal sample sequence X in the floating-point format, which is received from a selecting part **102**, to E−8, thereby shifting the decimal point toward the most significant bit position by 8 bits. An integer formatting part **12** truncates the fractional part following the decimal point having been shifted by 8 bits and adds “1” at the top of the resulting integer part to form a 16-bit (=24−8) signal sample sequence Y in the integer format. A compressing part **13** losslessly compresses the signal sample sequence Y and outputs a code sequence Ca.

An inverse digit adjusting part **32** shifts the signal sample sequence Y in the 16-bit integer format by 8 bits in the direction opposite to the shifting by the digit adjusting part **31**, that is, inserts 8 bits of “0” at the bottom of the 16-bit signal sample Y to shift the whole signal sample Y by 8 bits toward the most significant bit position, thereby forming a 24-bit signal sample in the integer format. A floating-point formatting part **15** converts the resulting signal sample in the 24-bit integer format into a signal sample sequence Y′ in the floating-point format. A subtraction part **16** determines the difference between the signal sample sequence Y′ in the floating-point format and the input signal sample sequence X, thereby producing a difference signal Z in the floating-point format. Actually, such processings by the inverse digit adjusting part **32**, the floating point formatting part **15** and the subtraction part **16** are not performed, and a difference producing part **14** can simply use the number of digits n of the integer part of the mantissa received from the integer formatting part **12** to extract the least significant 8+h (h=15−n) bits of the mantissa M_{X }of each input signal sample X in the case where n≠0, or to extract the exponent E and the mantissa M_{X }in the case where n=0. As in the case shown in **17** or **131** depending on the output of an analyzing part **132**.

The compressing part **17** comprises a divider **33** and two compressors **17** *h *and **17** *c*. The difference signal Z is input to the divider **33**. The distributor receives also the number of digits n of the integer part of the mantissa of each sample X after shifting from the integer formatting part **12**. In the case where n≠0, each difference signal sample Z is divided into a digit subset Zh containing bits capable of assuming non-zero determined by the number of digits n, that is, containing the least significant h bits (h=15−n) and a digit subset Zc containing bits capable of assuming non-zero determined by the conversion of the signal sample sequence Y from the 24-bit integer format into the 16-bit integer format or determined by the number of bits of the signal sample Y in the integer format, that is, containing the (h+1)th to (h+8)th least significant bits. In the case where n=0, each difference signal sample Z is divided into a digit subset Zh composed of 8 bits of the exponent E and the least significant (23−8) bits of the mantissa and a digit subset Zc composed of the most significant 8 bits of the mantissa. The compressors **17** *h *and **17** *c *losslessly compression-code the digit subsets Zh and Zc by entropy coding or the like and output code sequences Cb**1** and Cb**2**, respectively.

Therefore, the compressor **17** *h *primarily receives and losslessly compression-codes the least significant bit(s) of the mantissa M that can be non-zero. If the original digital signal is a 24-bit or 16-bit integer-value signal, the least significant h bits in the digit subset Zh input to the compressor **17** *h *are all “0” except the case where n=0 as shown in

As shown in **14**B and **14**C, the compressor **17** *c *losslessly compression-codes the 8 least significant bits that can be non-zero, that is, the (h+1)th to (h+8)th least significant bits, regardless of whether the input signal sample X is derived from a general floating-point signal, a 24-bit integer-value signal or a 16-bit integer-value signal. This enables improvement of the total compression efficiency, compared with the case where all the bits that can be non-zero in the difference signal Z are collectively compression-coded by the compressing part **17** shown in **17** *c *and **17** *h *can be applied not only to the coder according to the embodiment shown in **8**, **9** and **11**.

**203** outputs code sequences Ca, Cb**1** and Cb**2** according to an auxiliary code Cd received from an auxiliary code extracting part **202**, an expansion part **21** losslessly expansion-decodes the code sequence Ca to reproduce the signal sample sequence Y in the 16-bit integer format. An expansion part **23** comprises two expanders **23** *h *and **23** *c *and a combiner **41**. The code sequences Cb**1** and Cb**2** are losslessly expansion-decoded by the expanders **23** *h *and **23** *c*, respectively, and the resulting bit sequences Bh and Bc are input to the combiner **41**. The combiner **41** receives also the number of bits n following the most significant “1” in the integer part of each sample decoded by the expansion part **21**. In the case where n≠0, h (=15−n) bits are extracted from the bit sequence Bh output from the expander **23** *h*, 8 bits are extracted from the bit sequence Bc output from the expander **23** *c*, and the h bits and the 8 bits are combined to constitute the least significant h bits and the most significant 8 bits of the mantissa M, respectively. For the following samples, similarly, depending on the number of digits n, h bits are extracted from the bit sequence Bh, 8 bits are extracted from the bit sequence Bc, and the extracted bits are combined. In the case where n=0, 8 bits as the exponent E and 15 (=23−8) bits are extracted from the bit sequence Bh, and 8 bits are extracted form the bit sequence Bc. The 8 bits extracted from the bit sequence Bc are inserted at the top of the 15 bits from the bit sequence Bh, thereby forming the 23-bit mantissa.

A digit adjusting part **42** inserts 8 bits of “0” at the bottom of the decoded signal sample Y in the 16-bit integer format to shift the signal sample Y by 8 bits toward the most significant bit position, thereby producing a signal sample in the 24-bit integer format. Then, a floating point formatting part **15** converts the signal sample into a signal sample sequence Y′ in the 32-bit floating-point format. A combining part **24** combines the signal sample sequence Y′ and the difference signal Z to reproduce a signal sample sequence X in the floating-point format. In the case where the separating part **203** outputs a code sequence Cc according to the auxiliary code Cd, an expansion part **221** decodes the code sequence Cc, thereby directly reproducing the dignal sample sequence X.

In the above description, the signal sample sequence Y in the integer format is a 16-bit sample or a 24-bit sample. However, the number of bits contained in the signal sample Y may be any number m equal to or less than 24, and depending on the value of m, the number of digits capable of assuming non-zero in each signal sample Y in the integer format can be determined. Furthermore, in the example described above, a sequence of input signal samples X in the floating-point format may be divided into frames of 1024 samples, for example, or into blocks of consecutive samples, and compression coding may be performed for each of such divisional units. In the case where compression coding is performed on a divisional-unit basis, the number of bits converted into the integer format may be suitably determined for each divisional unit, and an auxiliary code that specifies that number of bits may be output.

As the compressing part **17** in the coder shown in **17** in the first embodiment shown in **23** in the decoder shown in **23** shown in **23** in the decoder decodes the code sequence Cb to reproduce the h bits, and inserts n bits of “0” at the top of the h bits to produce the mantissa.

In the embodiments described above with reference to **8**, **9**, **11** and **12**, for example, the compressing part **17** may be configured to perform for each frame the lossless compression coding on bit sequences extracted from the difference signal Z along the time-axis direction. In the following, such a coding method will be described with reference to

_{0}) and a bit sequence of the h-bit fractional part thereof (h=23−n). As described in the first embodiment with reference to _{j}, and the bit sequence E_{j }is coded.

The bit length h (=23−n) of the fractional part varies with the sample. Thus, in scanning the bits at the bit position j(j=0 to h_{max}−1; h_{max }represents the maximum bit length of the fractional part of the samples in the relevant frame) along the amplitude direction in the time-axis direction, if a sample that has no bit at the bit position j is encountered, the sample is skipped. Whether to skip a sample i or not can be determined as follows: at the bit position j along the amplitude direction of the sample i, the bit length h of the fractional part of the sample is compared with the value j, and if h≧j, the bit at the bit position is extracted, and if h<j, the sample is skipped, and the same determination concerning the next sample (i+1) is made.

For example, in the scanning, in the time-axis direction, of bits at the bit position j=17 from the MSB, the sample i=0, whose bit length h of the fractional part is 23−6=17, has the least significant bit at the bit position j. However, the sample i=1, whose bit length of the fractional part is 23−8=15, has no bits following the 15th MSB (or bit position j=14). Thus, the sample i=1 is skipped. The extracted 15th bits are grouped for the frame to form a bit sequence E**16**, and the bit sequence E**16** is entropy-coded. The same process is performed for each of the other bit positions j. As the entropy coding, arithmetic coding or Golomb coding is advantageous because layering can be provided.

On the decoder side, the number of digits n of the integer part can be determined from the number of bits n following the most significant “1” in each decoded integer-value signal sample Y, and therefore, the bit length h (=23−n) of the fractional part of the mantissa of each difference signal Z in the frame can be determined. Thus, for each of the decoded bit sequences E_{j}, the bits for the samples i in the frame are scanned sequentially, and the bit position j along the amplitude direction is compared with the bit length h for each sample. If j≦h, the fractional part of the mantissa shown in _{j }to the bit position (i, j). Then, at the top of the fractional part of each sample, n bits of “0” associated with the sample can be added, thereby reproducing the 23-bit mantissa M of the difference signal.

**17** shown in **8**, **9**, **11** and **12** to which the coding method shown in **17** receives one frame of samples z_{0}–z_{1023 }of the difference signal Z from the difference producing part **14** (see _{0}–n_{1023 }of their respective integer parts thereof in a storage part **17**B. Based on the number of digits n_{i }of each sample i, a control part **17**C places h-bit integer part (h=23−n) of the mantissa of the input difference signal Z_{i }excluding the most significant n bits in a buffer **17**A. Thus, the array of the integer parts of the mantissas as shown in **17**C scans, in the time-axis direction, the samples at the bit position j (=0 to h_{max}−1) along the amplitude direction, extracts the bits for which the relation of h≧j is satisfied to form a bit sequence E_{j}, and supplies the bit sequence E_{j }to a compressing part **17**D. The buffer **17**A, the storage part **17**B and the control part **17**C constitute bit sequence producing means. The compressing part **17**D losslessly compression-codes the bit sequence E_{j }and outputs a code Cb.

**23** shown in **10** and **13** serving as difference signal decoding means that is associated with the compression-coding part **17** shown in **23**D decodes the received code Cb to reproduce the bit sequence E_{j}. On the other hand, the numbers of digits n_{0}–n_{1023 }of the integer parts for one frame, which are received from the digit calculating part **21**A (see **23**B, and a control part **23**C allocates one bit extracted from the bit sequence E_{j }to each bit position (i, j) for which the relationship of h_{i}≧j concerning the number of digits h_{i }(=23−n_{i}) of the fractional part of each sample i is satisfied, and stores the resulting bit arrangement in a buffer **23**A, and repeats the same process for all the samples, thereby reproducing the fractional-part bit array shown in _{j }(j=0 to h_{max}−1) for one frame is completed, the control part **23**C adds n_{i }bits of “0” at the top of the fractional part of each sample i in the buffer **23**A to reproduce the mantissa of the difference signal, and supplies the difference signal Z to the combining part **24** (**23**A, the storage part **23**B and the control part **23**C constitute reconstructing means.

In the coding of the mantissa of the difference signal shown in _{j}, and the bit sequence E_{j }is coded. In the following, another method of scanning bit positions will be described. In _{2′}, E_{18′} and so on. Thus, the number of bit sequences E_{j }is less than the maximum number of bits h_{max }of the fractional part. In this bit scanning, again, any bit position for which the relationship h_{i}<j is satisfied is skipped. In the case of this bit scanning method, again, the number of digits of the fractional part of each sample i can be determined according to h_{i}=23−n_{i}. Thus, on the decoder side, the bit arrangement of the fractional part can be reproduced by allocating the bits in the bits sequences E_{j }in the same manner as described above.

_{i}<j is satisfied is skipped.

**0**, that is, in the case where the value of E−E_{0 }is negative, all the 32 bits including the exponent have to be coded and transmitted. Thus, as shown in **17** shown in

In the case where the input signal sample sequence X is a signal sample sequence in the floating-point format that is converted from the integer format, and the number of digits of the original integer value differs from the number of digits of the truncated integer value, that is, in the case where a signal sample sequence in the floating-point format directly converted from a 24-bit integer-value signal sample sequence is coded into a 16-bit integer-value signal sample sequence through truncation as described with reference to _{0 }to E_{7 }can be non-zero, although the following bit sequences E_{8 }to E_{21 }are all “0”, so that efficient coding can be achieved. It will be readily understood that efficient coding can be achieved similarly if the methods shown in

As is performed in the audio signal processing as required, for example, if a 24-bit integer-value signal sample is increased 1.5-fold (or 1.1-fold in binary notation), the resulting signal sample is converted into a signal in the floating-point format, and the resulting signal is coded into a 16-bit integer-value signal through truncation, increasing the 24-bit integer-value signal sample 1.5-fold makes at least the first decimal place be a bit that can be non-zero. Specifically, the first place following the decimal point P_{D }in the mantissa M_{X }of the sample X shown in _{Z }of the difference signal that can be non-zero increases toward the least significant bit position by 1 bit to 9 bits. In this case also, the bits following the 9 bits are all “0”, so that efficient coding can be achieved using the methods shown in **21**, **22** and **23**. Besides, if all the bits are not 0, the frequency of occurrence of non-zero (that is, “1”) may vary with the digit place, and in such a case, the lossless coding can be applied suitably for the variation.

In any case, the numbers of digits capable of assuming non-zero in the fractional part of the mantissa of the difference signals counted from the side of the MSB are equal, so that efficient compression coding can be achieved.

The mantissa coding methods shown in **17** shown in **23** shown in

In addition, the samples can be rearranged for scanning based on the number of digits of the fractional part. For example, the samples whose fractional part is composed of 23 digits are grouped together, the samples whose fractional part is composed of 22 digits are then grouped together, and samples having fractional parts composed of less numbers of digits are grouped together in the same way. The reverse arrangement of the samples can be achieved in a unique manner without the need of any additional information, so that the number of samples to be scanned can be simply managed.

In the example shown in _{j }is produced by scanning, in the time-axis direction, bits capable of assuming non-zero in the mantissa M of the samples z of difference signal Z in each frame, each sample containing h bits capable of assuming non-zero, with the MSB of the h bits of each sample aligned with the MSBs of the other samples. However, as shown in **17** can be configured the same as that shown in **23** in the decoder can be configured the same as that shown in

**17** in the coder shown in **8**, **9**, **11** and **12** described above. In this embodiment, an integer-value sample sequence is formed by arranging the fractional parts containing bits capable of assuming non-zero with the LSBs thereof aligned with each other, and compression coding is performed by eliminating the redundancy in the sample-sequence direction. As shown in **17** comprises a fractional-part extracting part **17**E, an in-frame averaging part **17**F, a subtracter **17**G and coding parts **17**D**1** and **17**D**2**. The fractional-part extracting part **17**E extracts, as one integer-value sample F, the fractional part composed of the least significant h bits (h=23−n) of the mantissa M of each difference signal sample Z received from the difference producing part **14**. The in-frame averaging part **17**F calculates the average F_{ave }of the extracted integer-value samples for each frame, and the coding part **17**D**1** codes the average F_{ave }and outputs the resulting code sequence Cb**1**. The subtracter **17**G determines the difference between the average F_{ave }and each integer-value sample F as an error sample ΔF, and the coding part **17**D**2** predictive-codes the error sample ΔF, losslessly compression-codes a prediction error and outputs the resulting code sequence Cb**2**. The predictive parameter may be separately transmitted as auxiliary information, or may be adaptively updated with the consistency being kept between the coder and the decoder.

**23** in a decoder associated with the coder shown in **23** in the decoder shown in **10** or **13**. According to this embodiment, the expansion part **23** comprises expanders **23**B and **23**C, an adder **23**D and an assembly part **23**A. The expander **23**B expansion-decodes the received code sequence Cb**1** and outputs the average F_{ave }for each frame. The expander **23**C expansion-decodes the received code sequence Cb**2** to determine the prediction error, performs predictive decoding and outputs the error sample ΔF. The adder **23**D sums the average F_{ave }and the error sample AF to reproduce the integer-value sample F. The assembly part **23**A assembles the difference signal Z from the integer-value sample F using the number of digits n from the digit calculating part **21**A.

This embodiment can be applied to the embodiment shown in **17** shown in **23** shown in **17** shown in **23** shown in

**17** shown in **8**, **9**, **11** and **12**.

As in the case shown in **17**B stores the numbers of digits n (n_{0}, n_{1}, n_{2}, . . . , n_{1023}) for difference signal samples received from an integer formatting part **12** (see **17**C, the least significant h (=23−n) bits, which can be non-zero in the mantissas M of the difference signals Z are stored in buffers **17**A**1**, **17**A**2** and **17**A**3**. In each of the buffers, the bits are stored with the MSB of each difference signal Z aligned with the MSBs of the other difference signals Z, as shown in **17**C, the bits in the bit array stored in the buffer **17**A**1** are scanned in the time-axis direction for each bit position along the amplitude direction to produce a bit sequence E_{j} **1**, as shown in **17**A**2** are scanned in the time-axis direction within each of predetermined bit lengths along the amplitude direction to produce a bit sequence E_{j} **2**, as shown in **17**A**3** are scanned in the in the amplitude direction within each predetermined bit length along the amplitude direction of each sample to produce a bit sequence E_{j} **3**.

For each frame, the control part **17**C determines the deviation of values of bits in each of the bit sequences Ej produced by the scanning methods in terms of count of “1”s for example, determines the degree of deviation of bits “1” for each scanning method, determines which of the scanning methods provides the highest degree of deviation (or the lowest entropy), and then supplies, to a selecting part **17**E, a selection signal SL that designates the bit sequence E_{j }in the buffer corresponding to the scanning method of the highest degree of deviation. The selected bit sequence is coded in a coding part **17**D, and the resulting code sequence Cb is output.

The entropy corresponding to the deviation of bits “1” in a frame is represented by the following formula:

Σ_{j} *−M* _{j}((*k* _{j} */m* _{j})log_{2}(*k* _{j} */m* _{j})+(1−(*k* _{j} */m* _{j}))log_{2}(1−(*k* _{j} */m* _{j}))) (2)

where “m_{j}” denotes the number of bits in each bit sequence E_{j}, and “k” denotes the number of “1”s in the bit sequence. “Σ_{1}” means a sum for all the possible “j” values (that is, the number of bit sequences). As the deviation of bits “1” increases, the entropy decreases. The entropy depends on how the bit array for the frame is scanned or how the bit array is divided into bit sequences. Thus, according to the seventh embodiment, the optimal scanning method for the bits capable of assuming non-zero is selected for each frame.

**23** shown in **10** and **13** that is adapted for the compressing part **17** shown in **23** is configured the same as the expansion part **23** shown in **23**E is additionally provided. The decoding part **23**E decodes an auxiliary code Ce to reproduce the selection signal SL and supplies the selection signal SL to a control part **23**C. Based on the scanning method and the number of digits that correspond to the supplied control signal, the control part **23**C stores the bit sequence E_{j }output from a decoding part **23**D into a buffer **23**A in such a manner that the sequences of bits capable of assuming non-zero for the respective samples are reproduced. Then, n_{i }bits of “0” are added at the top of the h_{j }bits capable of assuming non-zero of each sample to form the mantissa M, and the mantissa M is output as the difference signal Z.

_{j} **1**, E_{j} **2** and E_{j} **3** from buffers **17**A**1**, **17**A**2** and **17**A**3** are coded by coding parts **17**D**1**, **17**D**2** and **17**D**3**, respectively. An analysis part **17**G determines which of the output codes is of the smallest quantity of information, and supplies a selection signal SL designating the output code not only to a selecting part **17**F but also to a coding part **17**E to obtain a code Ce. The selecting part **17**E outputs the code sequence of the smallest quantity of information designated by the selection signal as a code sequence Cb, along with the code Ce. As an expansion part **23** in the decoder associated with the compressing part **17** shown in **23** shown in

Instead of the compressing parts **17** and **131** shown in **17**-**1**, a variable bit extracting part **17**-**2**, a mantissa compressing part **131**-**1** and a mantissa extracting part **131**-**2**. The variable bit compressing part **17**-**1** extracts and compression-codes h (=23−n) bits (variable bits) capable of assuming non-zero from the mantissa M of the difference signal Z based on the number of digits n and outputs a code Cb. The variable bit extracting part **17**-**2** extracts a bit sequence M_{D }containing h bits capable of assuming non-zero from the mantissa M of the difference signal Z and outputs the bit sequence without processing. The mantissa extracting part **131**-**1** compression-codes the mantissa M of the difference signal Z and outputs a code Cm. The mantissa extracting part **131**-**2** extracts the mantissa M of the difference signal Z and outputs the mantissa M without processing. An efficiency estimating part **134** calculates the quantity of information of each of Cb, M_{D}, Cm, M and Z, and determines one of them which is of the smallest quantity of information. The selection signal SL designating the determined one is coded by an auxiliary coding part **103**, and the resulting code Ce is supplied to an output part **104**. The output part **104** selects one of Cb, M_{D}, Cm, M and Z which is designated by the selection signal SL and outputs the one as the code Cb along with the code Ce.

**23** and **231**, the selecting part **233** and the auxiliary decoding part **234** in the decoder shown in **232** is not necessary. In **234** decodes the code Ce into the selection signal SL, and based on the selection signal SL, a selecting part **233** is controlled. Specifically, depending on the selection signal SL, as the input code Cb, the compression code Cb of the variable bit sequence is supplied to a decoding part **23**, the variable bit sequence (the sequence of bits capable of assuming non-zero) M_{D }is supplied to an assembly part **23** *a*, the compression code Cm of the mantissa M is supplied to a decoding part **231**, or the mantissa M is output as it is. The decoding part **23** decodes the code sequence Cb to produce the bit sequence M_{D }composed of bits capable of assuming non-zero and supplies the bit sequence M_{D }to the assembly part **23** *a*. Based on the number of digits n, the assembly part **23** *a *adds n bits of “0” at the top of the bit sequence M_{D }to reproduce the mantissa M.

In **15** is made equal to the exponent of the corresponding sample in the input floating-point signal sample sequence X, rounding can be used instead of truncation. As shown in

Thus, rounding is modified as described below. As shown in _{I }of the mantissa M assumes a value 2^{n}−1, which precedes a power of 2, truncation, rather than rounding, is performed. Otherwise, rounding is performed. Although 1 bit is additionally required to indicates the sign that designates whether the error signal (difference signal) Z is positive or negative, the maximum number of digits of the mantissa is reduced by 1 bit, because the error of the rounded sample is equal to or less than 0.5. Thus, the number of digits of the mantissa including the sign bit is the same as in the case of truncation. However, when the integer part M_{I }assumes a value of 2^{n}, that is, a power of 2, the number of digits of the mantissa of the error can be reduced by 1, and when the integer part M_{I }assumes a value of 2^{n}−1, an error up to 1 occurs only if the error is positive. Thus, these two cases have to be handled specially.

That is, the sign bit is added at the top of the bits capable of assuming non-zero in each difference signal sample Z, a code designating the number of bits h capable of assuming non-zero is added to the bits and the resulting bit sequence is compression-coded if the integer part M_{I }assumes a value of 2^{n}−1, and the decoder decodes and expands the code sequence Cb and separates the resulting bit sequence into sub-sequences each containing a number of bits, the number of bits being a sum of the number of bits h capable of assuming non-zero determined from the number of bits n of the integer part and one bit of the sign bit. If the integer part M_{I }assumes a value 2^{n}−1, the sign bit, the number of bits of the code designating the number of bits h capable of assuming non-zero, and the number of bits h designated by the code are separated to determine one sample of difference signal Z. In the case where only the bits capable of assuming non-zero are coded, the efficiency increases as the amplitude increases. As the amplitude increases, the number of times that the integer part assumes a value of 2^{n}−1, that is, the number of times of occurrence of the exceptional processing decreases, so that the compression efficiency can be improved in total.

Using the modified rounding described above can make the error energy between the floating-point input and the integer part lower than that in the case of using truncation, and thus, the prediction efficiency for the integer part sequence can be improved. In addition, because only the bits capable of assuming non-zero have to be coded, the efficiency of compression of the difference signal Z by the compressing part **17** can be improved.

For example, as shown by an alternate long and short dash line in **2**, it is determined whether the integer part M_{I }of the input floating-point signal sample sequence X is 2^{n}−1 or not. For example, it is determined whether the most significant n (=E−E_{0}) bits of the mantissa M are all “1” or not. If all the bits are “1”, it means that the integer part M_{I }assumes a value preceding a power of 2, the process continues to step S**3**, where truncation is performed. If the most significant E−E_{0 }bits are not all “1”, the process continues to step S**6**, where the integer formatting through the modified rounding is performed.

In the second embodiment shown in **132**, a selection is made between compression coding by the compressing part **17** and compression coding by the compressing part **131**. To achieve proper selection, as shown in **17** and **131** perform coding, and one of the resulting codes which is of higher compression efficiency is adopted. Although the compression coding by the compressing part **131** is accurate, the direct compression coding of the difference signal Z requires a relatively large processing quantity and processing time. Thus, the method described below may be used.

For example, as a method of estimating the efficiency of universal coding on a byte (8-bit) basis, there is a method of calculating the entropy for each byte. An error signal, that is, a difference signal Z in the floating-point format is regarded as an integer value y ranging from 0 to 2^{8}−1 (=255) on an 8-bit basis, a histogram is created from the number of times of occurrence L(y) of the integer value y within one frame, and the number of times L(y) is divided by the number of bytes B in the frame to determine a density function p(y).

*p*(*y*)=*L*(*y*)/*B* (3)

From this density function p(y), an average entropy H(p(y)) is determined.

The universal coding can possibly compress 1 byte of numeral value to about the size of the average entropy (bit). Thus, if the product of the average entropy value and the number of bytes B in the frame is smaller than the total number of bits in one frame resulting from compression coding of only the bits capable of assuming non-zero by the compressing part **17**, the universal coding on a byte basis, that is, the compression coding by the compressing part **131**, is more advantageous.

A configuration of a coder that allows for such selection between compressing parts is shown in **133**, an average entropy calculating part **136** and a determination part **137**. As shown in this drawing, a difference signal Z is input from a difference producing part **14** to the average entropy calculating part **136**, the average entropy H(p(y)) is determined as described above, and the product of the average entropy H(p(y)) and the number of bytes B is determined. The determination part **137** compares the product with the number of bits of one frame of code sequence Cb′ estimated by an efficiency estimating part **134** *a*. If the former is smaller, the selecting part **133** is controlled so that the difference signal Z is compression-coded by the compressing part **131**, or if the latter is smaller, the code sequence Cb′ produced by the compressing part **17** through compression coding is output as the code sequence Cb. Also, an auxiliary code Ce indicating which compressing part is selected is outputted. Here, if an integer formatting part **12** performs integer formatting by truncation in the calculation of the entropy, the number of times of occurrence L(y) of the 23-bit mantissa M of each difference signal sample Z is determined on a 1-byte basis, or if the integer formatting part **12** performs integer formatting by the modified rounding, the number of times of occurrence of each sample including the sign thereof is determined.

If the efficiency of the universal coding is estimated as described above, the processing load can be reduced compared to the case where the difference signal Z is universal-coded, the total number of bits of the resulting code within one frame is determined, and the analyzing and selecting part **134** selects one of the compressing parts. The use of such an efficiency estimation of the universal coding can be applied to the embodiment shown in

In the above description, the samples in the signal sample sequence Y in the integer format have been 16-bit samples or 24-bit samples. However, the samples in the signal sample sequence Y may be composed of any number of bits equal to or less than 24, and based on the number of bits, the number of bits capable of assuming non-zero in each sample in the digital signal sample sequence Y in the integer format, which depends on the number of bits of each sample, can be determined. In this way, if the compression coding is performed on a divisional-unit basis, the number of bits to be converted into the integer format can be conveniently determined, and an auxiliary code that designates the number of bits can also be output.

[Implementation by Computer]

As shown in **8**, **9**, **11**, **12**, **16** and **35** can be implemented by a computer **600** comprising a RAM **61**, a CPU **62**, a hard disk **63**, a ROM **64**, a transmitter/receiver **65**, an input/output part **66** and an audio/video input/output part **67**, which are interconnected via a bus **68**.

The ROM **64** stores a program for booting the computer, and the hard disk **63** stores an operating system program for the computer. A program for executing a function of the coder according to the present invention is previously read into the hard disk **63** from a recording medium **71**, such as CD-ROM or DVD, via the CD/DVD drive **66**, or previously downloaded to the hard disk **63** through a communication line and the transmitter/receiver **65**.

The signal sample sequence X in the floating-point format to be coded is externally received at the audio/video input/output part **67**, for example, and temporarily stored in the hard disk **63**, which serves as a buffer. To start coding, a program for executing coding is read into the RAM **61** from the hard disk **63**, and the program is executed by the CPU **62**. The result of coding, such as the codes Ca, Cb (or Z) and Cc, may be transmitted from the transmitter/receiver **65** to the outside via the communication line, or may be stored in a recording medium **65**, such as CD and DVD, in the CD/DVD drive **66**.

Similarly, in order to implement the decoder shown in **10**, **11**, **13** and **17** by the computer shown in **63**, and a decoding program is executed on the codes Ca, Cb (or Z) and Cc received at the transmitter/receiver **65** through the communication line. The result of coding is output to a display unit **73** and a speaker **74** for reproduction.

The present invention can be applied not only to music signals but also to sound signals, image signals and the like.

Patent Citations

Cited Patent | Filing date | Publication date | Applicant | Title |
---|---|---|---|---|

US6633679 * | Mar 22, 2000 | Oct 14, 2003 | Teletechnologies, Ltd. | Visually lossless still image compression for CMYK, CMY and Postscript formats |

US20050091051 * | Mar 10, 2003 | Apr 28, 2005 | Nippon Telegraph And Telephone Corporation | Digital signal encoding method, decoding method, encoding device, decoding device, digital signal encoding program, and decoding program |

JP2001044847A | Title not available | |||

JP2003332914A | Title not available | |||

WO2003077425A1 | Mar 10, 2003 | Sep 18, 2003 | Nippon Telegraph & Telephone | Digital signal encoding method, decoding method, encoding device, decoding device, digital signal encoding program, and decoding program |

Non-Patent Citations

Reference | ||
---|---|---|

1 | Hans, Mat et al., "Lossless Compression of Digital Audio", IEEE Signal Processing Magazine, pp. 21-32, 2001, no month. | |

2 | Jin, Akio, et al., A Hierarchical Lossless/Lossy Coding System for High Quality Audio up to 192 KHz Sampling 24 Bit Format, IEEE Transactions on Consumer Electronics, vol. 49, No. 3, pp. 759 to 764, 2003, no month. | |

3 | Moriya, Takehiro et al., "Onkyo Shingo no Sampling Rate Scalable Lossless Fugoka", Joho Kagaku Gijutsu Forum Ippan Koen Ronbunshu, Separate vol. 2, pp. 227 to 228, 2002. (with English Translation), no month. |

Referenced by

Citing Patent | Filing date | Publication date | Applicant | Title |
---|---|---|---|---|

US7408481 * | Oct 14, 2005 | Aug 5, 2008 | Nippon Telegraph And Telephone Corporation | Information encoding method, decoding method, common multiplier estimating method, and apparatus, program, and recording medium using these methods |

US7436969 * | Sep 2, 2004 | Oct 14, 2008 | Hewlett-Packard Development Company, L.P. | Method and system for optimizing denoising parameters using compressibility |

US7605722 * | Apr 11, 2008 | Oct 20, 2009 | Electronics And Telecommunications Research Institute | Lossless audio coding/decoding apparatus and method |

US8571112 * | May 12, 2010 | Oct 29, 2013 | Huawei Technologies Co., Ltd. | Specification method and apparatus for coding and decoding |

US8947274 * | Jun 21, 2012 | Feb 3, 2015 | Mitsubishi Electric Corporation | Encoding apparatus, decoding apparatus, encoding method, encoding program, decoding method, and decoding program |

US8959129 | Jun 27, 2012 | Feb 17, 2015 | Altera Corporation | Compression of floating-point data |

US8959130 | Jun 27, 2012 | Feb 17, 2015 | Altera Corporation | Compression of floating-point data in encoding groups |

US9047118 | Jun 27, 2012 | Jun 2, 2015 | Altera Corporation | Computationally efficient compression of floating-point data |

US9104473 | Sep 14, 2012 | Aug 11, 2015 | Altera Corporation | Conversion and compression of floating-point and integer data |

US20060047484 * | Sep 2, 2004 | Mar 2, 2006 | Gadiel Seroussi | Method and system for optimizing denoising parameters using compressibility |

US20100290535 * | Nov 18, 2010 | Fengyan Qi | Specification method and apparatus for coding and decoding | |

US20140313064 * | Jun 21, 2012 | Oct 23, 2014 | Mitsubishi Electric Corporation | Encoding apparatus, decoding apparatus, encoding method, encoding program, decoding method, and decoding program |

Classifications

U.S. Classification | 341/51, 704/229, 704/E19.004 |

International Classification | H04N19/60, H04N19/70, H04N19/12, H04N19/91, H04N19/146, H04N19/196, H03M7/36, H04N19/00, G10L19/00, H03M7/30, H03M7/38, H03M7/24 |

Cooperative Classification | H03M7/24, H03M7/30, G10L19/0017 |

European Classification | G10L19/00L, H03M7/24, H03M7/30 |

Legal Events

Date | Code | Event | Description |
---|---|---|---|

Mar 9, 2007 | AS | Assignment | Owner name: NIPPON TELEGRAPH AND TELEPHONE CORPORATION, JAPAN Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:MORIYA, TAKEHIRO;YANG, DAI;HARADA, NOBORU;REEL/FRAME:018987/0989 Effective date: 20051028 |

Aug 19, 2010 | FPAY | Fee payment | Year of fee payment: 4 |

Aug 1, 2014 | FPAY | Fee payment | Year of fee payment: 8 |

Rotate